ENSO Forecasting from Sea Surface Temperature Observations via Structured Variational Autoencoders

The El Niño-Southern Oscillation (ENSO) is the dominant driver of year-to-year climate variability, and seasonal ENSO forecasts inform agricultural planning, water management, and anticipatory humanitarian action worldwide. We forecast the Oceanic Niño Index (ONI) with a structured variational autoencoder (SVAE): a convolutional encoder-decoder coupled to a month-conditioned linear latent dynamical system, trained end-to-end only on observed sea-surface temperature (SST) (no climate-model simulations or other predictors). Under a strictly leakage-free hindcast protocol (2002-2022), the SVAE outperforms established statistical and deep-learning baselines in root-mean-square error (RMSE) and anomaly correlation coefficient (ACC) across most lead times. Crucially, we show that the model \emph{learns} to embed the ONI index along the leading latent direction while reconstructing spatial SST maps, even when trained without a dedicated Niño3.4 loss term.